The unique properties of a memristor make it promising in neuromorphic computing systems: First, the prototyped memristor devices have demonstrated the scalability down to 10 nm and below. Accordingly, the memristor memories can achieve a high integration density of 100Gbits/cm2, a few times higher than the flash memory technologies. Second, the memristor device has an intrinsic and remarkable feature called “pinched hysteresis loop”, which means it can “remember” the total electric charge flowing through it. Consequently, the memristor-based memories with both the high-integration capability and the pinched hysteresis characteristic can be applied to a massively-parallel, large-scale neuromorphic computing processor architecture.
In this project, we investigated memristor-based reconfigurable design techniques. The structure is built upon single memristor-based synapse and the corresponding training circuit design. The design optimization and its implementation in multi-synapse systems are also analyzed. With the aid of sharing training circuit and self-training mode, the performance and energy can be significantly improved. The proposed techniques can potentially be extended to develop a neuromorphic computing processor architecture.
Senior Personnel: | Prof. Hai (Helen) Li (PI) |
Sponsors: | US Air Force Research Lab |
Team Members: | Miao Hu and Hui Wang |
Collaborators: | Dr. Qing Wu, U.S. Air Force Research Lab |
Performance Date: | 10/01/2010 – 09/30/2012 |
Related Publications:
[2012]
H. Li and R. E. Pino, “Statistical Memristor Model and Its Applications in Neuromorphic Computing,” (in Advances in Neuromorphic Memristor Science and Applications, Editor: R. Kozma, R. E. Pino, and G. Pazienza), Springer, Jun. 28, 2012. ISBN: 978-94-007-4490-5.
[2011]